Advanced Networking Technologies

Transcription

1 Advanced Networking Technologies Chapter 13 Caching Techniques for Streaming Media (Acknowledgement: These slides have been prepared by Dr.-Ing. Markus Hofmann) 1 What is Streaming? Streaming media refers to media types with time constraints and a continuous data flow. Video and audio transmission. Playback of streaming media starts while data is being received, i.e. it is not necessary to download the entire file before playback starts Different from MP3 downloads (e.g. using Napster). Streaming media is becoming more and more important on the Internet. Commercial usage, Private usage. The streaming media server, advertising, subscription services, and online music shipment market is expected to grow to $14.9 billion by (Source: Web page of Global Information, Inc.) 2

2 Application Examples Live, non-interactive applications: Internet radio, news broadcasts, sporting events, etc., Content is typically not recorded in advance, Loose delay constraints. Live, interactive applications: Teleconferencing, video/audio phones, distributed games, etc., Tight delay constraints to support interactivity. Stored, non-interactive applications (on-demand): Movies over the Internet, news archives, video clips on homepages, etc., Content recorded in advance; loose delay constraints. 3 Application Software Public domain software for streaming over the MBone: VideoConference (VIC), INRIA Video Conferencing System (IVS), Visual Audio Tool (VAT), Robust Audio Tool (RAT), Network Voice Terminal (Nevot). Dominant players in the commercial world: Microsoft (Windows Media Player, Silverlight), Adobe (Flash), Apple (QuickTime Player). 4

3 Example: RealPlayer ( click play button ) 5 Example: RealPlayer ( buffering ) 6

4 Example: RealPlayer ( playing ) 7 Streaming on the Internet - Today Current streaming technology is in a relatively primitive stage, despite great research work. Dominant market players are Microsoft, Adobe, and Apple Most streaming applications are based on unicast transport: Server load increases linearly with the number of clients. Bandwidth intensive multimedia flows lead to serious network congestion. Clients experience unpredictable playback quality and high start-up latency. Origin Server Internet Clients 8

5 Example: Clinton s Testimony CNN audio and video quality became unbearable for most people at around 1:00 pm on August 17, Link to video stream removed from CNN Web page by 1:15 pm. Other news servers were also unreachable. 9 Using IP Multicast for Streaming Media IP multicast replicates data packets at branching points in the network. Benefits of using IP multicast technology: Reduction of server load, Reduction of network load. Problems with multicast-based streaming: IP multicast is not yet widely deployed (lack of business model, management issues, does not yet work properly across domains, etc.). Multicast is useful for live broadcasts; it is not directly applicable to on-demand style services (because it requires synchronous receivers). Multicast does not address the problem of high start-up latency. Origin Server Internet Clients 10

6 Caching of Streaming Media Client requests for streaming media are handled by the local cache: Cache miss: Request is forwarded to the origin server, which starts playback through the cache; the cache relays the data and simultaneously stores the stream. Cache hit: The cache starts playback of the requested stream. Benefits of caching: Reduced server load, Reduced network load, Improved playback quality, Improved start-up latency. Origin Server Internet Cache Clients 11 Which Would You Rather Be Watching? A loading bar... or the real image 12

7 13 Buffering Connecting Playing sec 14

8 Server no longer reachable Please try again later Network Congestion

9 Connecting Buffering Playing The Answer is easy The real image!!! 18

10 Caching Techniques for Streaming Media Why is streaming media different from conventional web traffic? Different signaling protocols (HTTP vs. RTSP), Object size: caching of streaming objects in their entirety does not scale, Time constraints of continuous media. Classical caching techniques are not feasible for streaming media. Innovative streaming solutions will adapt to user behavior and will scale with increasing demand of streaming media. Fast prefix transfer reduces access delay, Scalability is achieved by segmentation of streaming objects, Dynamic caching increases throughput through request aggregation. 19 Streaming Media Protocols IETF Standard Streaming Protocols RTSP: Real Time Streaming Protocol, RFC Control protocol for streaming (like a VCR remote control), Superset of HTTP 1.1 with streaming extensions. RTP: A Transport Protocol for Real-Time Applications, RFC Framework for transport of real-time data. RTCP: The RTP Control Protocol. Enables feedback on session quality/status. DASH Widely-Used Proprietary Protocols PNA: Progressive Network Architecture. Real Networks proprietary protocol corresponding to RTSP. RDT: Real Data Transport protocol. Real Networks proprietary protocol corresponding to RTP. MMS: Microsoft s proprietary protocol for streaming. 20

11 Real Time Streaming Protocol (RTSP) RTSP controls transmission of streaming objects ( control for remote VCRs ). RTSP is different from HTTP. Data carried out of band (RTP), RTSP server is state full, Uses different methods (PLAY, PAUSE, RECORD, etc.). RTSP inherits from HTTP. Re-use HTTP extensions. n Authentication, n Cache control. Media stream and sessions defined by a RTSP URL. rtsp:// 21 Layering Of Streaming Protocols Application Layer HTTP RTSP RTP RTCP Transport Layer TCP UDP Network Layer IP 22

12 RTSP Operations HTTP GET Session description Web Server SETUP RTSP Client PLAY RTP audio RTP video RTCP RTSP Server PAUSE TEARDOWN 23 On The Need Of Client Side Buffering (1) Server Client Playback Interruption 24

13 On The Need Of Client Side Buffering (2) Server Client Buffer 25 Buffered Playback Server Client Client Request Connection Delay Buffer Delay Playback Streaming applications maintain a playback buffer to absorb jitter. Playback starts when the playback buffer is filled completely. 26

14 Example: Access Delay Access Delay [sec] Vintage PLJ FM Champla NDR 2 SRA-FM Rock Speicherverzögerung Buffer Delay [sec] [sek] Verbindungsverzögerung Connection Delay [sec] [sek] 27 Fast Prefix Transfer Cache Client Client Request Playback Fast prefix transfer at the beginning of the data transfer decreases buffer delay. Knowledge of the buffer size at the client allows for optimal improvement. 28

15 Segmentation of Streaming Objects (1) Without Segmentation With Segmentation A B A 1 B A 2 A 1 initial cache content cache content after receiving request for object B initial cache content cache content after receiving request for object B Division of streaming objects into smaller segments. Segments can be cached and replaced independently. 29 Segmentation of Streaming Objects (2) A 1 A 2 C 1 A S C R C 2 A 3 A 4 Cooperation among caches allows the use of segments from multiple caches Timing relation among segments must be controlled: Intelligent fetching of segments, Intelligent cache replacement to avoid gaps. 30

16 Cache Replacement for Streaming Media A/V-Object Chunk 0 Chunk 1 Chunk 2 Chunk 3 cached segments Non-cached segments Goal: fine grained cache replacement, while avoiding data gaps. Cache replacement algorithm: Determine the victim segment, Determine the chunk to which the victim belongs, Eject the last segment of the victim chunk. 31 Dynamic Caching S R 1 t Data for R 2 Patch C 1 Data for R 1 Δ R 2 t + Δ t t+δ t+2δ Asynchronous requests characterized by temporal distance Δ. Dynamic cache bridges temporal distance. Enables use of multicast even for asynchronous requests. Size of dynamic cache independent from size of media object. Dynamic caching requires data patching. 32

17 Simulation Results Network simulation based on MCI backbone topology using network simulator ns-2. Servers holds 100 different streaming objects. Access to streaming objects according to Zipf distribution. Playback rate: 1,33 Mbps. Size of streaming objects: 1 GB. Cache size: 5 GB. Simulation time: 10 hours. Time interval between two successive client requests: 10 sec. 33 Benefits of Segmentation # data streams from sender No Segmentation 167 MB/Segment 16,7 MB/Segment 1,67 MB/Segment Time 34